A dependency-free terminal coding agent.
Drinky can read, search, and change files or run commands in the working directory. The conversation
stays in the terminal scrollback. Drinky saves no transcript. /gist writes a compressed summary of
the current task, and Drinky saves it after a review. /gists carries a saved gist into a new
conversation. A gist leaves out tool calls and reasoning, so it usually works even after a context
overflow.
Drinky supports Anthropic, OpenAI, xAI, OpenRouter, DeepSeek, and Gemini Enterprise Agent Platform (Vertex AI). It learns models from the providers and uses public metadata for prices and model capabilities.
Drinky is a single Zig program. Instruction files and skills define how the agent works. The model
can read the command, config, and key reference through the describe_drinky tool.
Drinky requires Zig 0.17.0, a POSIX system, and the HOME variable. A terminal with the Kitty
keyboard protocol and grapheme cluster processing gives the best experience.
zig build -Doptimize=safe
./zig-out/bin/drinkyType / to list the commands. Ask Drinky about its commands, key bindings, or config.
| Account | Credential |
|---|---|
anthropic-plan |
Claude Pro or Max login |
anthropic-api |
Anthropic Console login, which mints and stores an API key |
anthropic-api-key |
ANTHROPIC_API_KEY |
openai-plan |
ChatGPT login |
openai-api-key |
OPENAI_API_KEY |
xai-plan |
SuperGrok or X Premium login with a device code |
xai-api-key |
XAI_API_KEY |
openrouter-api |
OpenRouter login, which mints and stores an API key |
openrouter-api-key |
OPENROUTER_API_KEY |
deepseek-api-key |
DEEPSEEK_API_KEY |
google-cloud-key |
GOOGLE_APPLICATION_CREDENTIALS and GOOGLE_CLOUD_LOCATION |
Use /login to sign in or switch accounts. Use /model to select a model.
Drinky is not affiliated with the providers.
Drinky reads each AGENTS.md file from the Git root down to the working directory. It finds skills
in ~/.agents/skills/ and in each .agents/skills/ directory on that path, and loads a skill on
demand.
The required_skills config key pairs a path pattern with a skill. Drinky sends the skill to the
model when a tool first touches a matching file. It refuses a write or an edit of that file until
the conversation holds the skill.
Use drinky run to answer a prompt without a terminal. A run reads the prompt from stdin and writes
the final reply to stdout. It uses the instructions and tools of a terminal session. The command
drinky models lists the values of the --model flag.
./zig-out/bin/drinky run --model openai-plan/gpt-5.5 --effort high <<'EOF'
Review the uncommitted changes.
EOFThe model can start a run through the bash tool, for example to review its work as a second agent. The session total then includes the available costs of the run.
A local MCP client can control running Drinky instances. For example, a ChatGPT voice chat can delegate project work to Drinky through the Codex runtime of the ChatGPT desktop app.
Type /mcp in an instance to turn MCP access to that instance on or off. Access starts off unless
the config file turns it on. The client starts drinky mcp as its stdio server. The server speaks
MCP 2026-07-28 and finds each instance with MCP access, so one server entry covers all instances.
The client can read the state, send messages, load skills, cancel work, clear the conversation, fetch a model list, and select the model and the effort. It cannot rewind the conversation, use gists, or sign in.
The optional ~/.drinky/config.json file controls instructions, required skills, request and
command limits, the default effort, and interface settings. The describe_drinky reference
documents each key, so Drinky can explain the file.
Drinky reports its state through the Program Status Protocol (OSC 7501) and to Herdr. A terminal such as Rex can then show when Drinky works, waits for the user, or has a result.
Drinky has no permission gate or sandbox. Use a container for an untrusted repository. MCP access lets any process of the same user submit work to the instance.
Drinky is available under the MIT License.